Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DALL·E now available in beta

openai.com

Post reply on HN
Print view View original

311–320 of 579 posts

Re: DALL·E now available in beta

#311

Post by slavak »

zanderwohl wrote: post #294
tiborsaas wrote: post #261

Earlier quoted context omitted.

Thanks, finally a legit news publication :) This was really funny :) http://dailywrong.com/man-finally-comfortable-just-holding-a...

Somehow these articles are more readable than typical AI-generated search engine fodder... Is it because I'm entering the site with an expectation of nonsense?

Probably because, by the creator's own admission, the articles are heavily cherry-picked to make sure the output is decent, which is probably a lot more human effort than goes into the aforementioned search engine fodder.

http://dailywrong.com/sample-page/

Re: DALL·E now available in beta

#312

Post by dave_sullivan »

ninjaranter wrote: post #310
dave_sullivan wrote: post #307

Earlier quoted context omitted.

Facebook released over 100 pages of notes a few months ago detailing their training process for a model that is similar in size. Does anyone have a link? I can't seem to find it in my notes, googling links to posts that have been removed or are behind the facebook walled garden. But I seem to remember they were running 1,000+ 32gb GPUs for 3 months to train it and keeping that infrastructure running day-to-day and tw…

> Facebook released over 100 pages of notes a few months ago detailing their training process for a model that is similar in size. Does anyone have a link? Is https://github.com/facebookresearch/metaseq/blob/main/projec... what you're referring to?

Yes! Thank you! Very good read for anyone interested in the field.

Re: DALL·E now available in beta

#313

Post by judge2020 »

binarymax wrote: post #271
TigeriusKirk wrote: post #252

Earlier quoted context omitted.

What are the resources at work here? What are the resources needed to train this model? If someone just gave you the model for free, what resources would you need to use it to generate new results?

If I had to guess, based on other large models, it’s in the range of hundreds of GBs. It might even be in the TB range. To host that model for fast production SaaS inference requires many GPUs. An A100 has 80GB, so a dozen A100s just to keep it in memory, and more if that doesn’t meet the request demand. Training requires even more GPUs, and I wouldn’t be surprised if they used more than 100 and trained over 3 months…

> Training requires even more GPUs, and I wouldn’t be surprised if they used more than 100 and trained over 3 months.

Based on this blog post where they scale to 7,500 'nodes', they say:

> A large machine learning job spans many nodes and runs most efficiently when it has access to all of the hardware resources on each node.

So I wouldn't be surprised if they do have a total of 7500+ GPUs to balance workloads between. TO add, OpenAI has a long history of getting unlimited access to Google's clusters of GPUs (nowadays they pay for it, though). When they were training 'OpenAI Five' to play Dota 2 at the highest level, they were using 256 P100 GPUs on GCP[0] and they casually threw 256 GPUs at 'clip' for a short while in January of 2021[1].

As for how they do it, see these posts:

https://openai.com/blog/techniques-for-training-large-neural...

https://openai.com/blog/triton/

0: https://openai.com/blog/openai-five/

1: https://openai.com/blog/clip/

Re: DALL·E now available in beta

#314

Post by gwern »

neya wrote: post #186

I'm curious to know - does the community have any open source alternatives to DALL.E? For an initiative named OpenAI, keeping their source code and models closed behind a license is bullshit in my opinion.

EAI/Emad/et al's 'Stable Diffusion' model will be coming out in the next month or so. I don't know if it will hit DALL-E 2 level but a lot of people will be using it based on the during-training samples they've been releasing on Twitter.

Re: DALL·E now available in beta

#316

Post by jnovek »

peteforde wrote: post #228

I have been having a blast with DALL-E, spending about an hour a day trying out wild combinations and cracking my friends up. I cannot imagine getting bored of it; it's like getting bored with visual stimulus, or art in general. In fact, I've been glad to have a 50/day limit, because it helps me contain my hyperfocus instincts. The information about new pricing is, to me as someone just enjoying making crazy imagines…

My heart sank when I saw the pricing model.

I’ve been creating generative art since 2016 and I’ve been anxiously waiting for my invite. I wont be able to afford to generate the volume of images it takes to get good ones at this price point.

I can afford $20/mo for something like this but I just can’t swing $200 to $300 it realistically takes to get interesting art out of these CLIP-centric models.

Heck, the initial 50 images isn’t even enough to get the hang of how the model behaves.

Re: DALL·E now available in beta

#317

Post by gwern »

can16358p wrote: post #219

So can we now legally remove the "color blocks" watermark or not? What about generating NFTs? It was explicitly prohibited during the previous period, now there is no notion of it. Without notion and rights for commercial use I think it's allowed but because it was an explicitly forbidden use case before, I want to be sure whether it can be used or not. Regardless, excited to see what possibilities it opens.

Another user saying that OA has said it's OK to remove the watermark: https://www.reddit.com/r/dalle2/comments/w3qsxd/dalle_now_av...

The commercial use language appears pretty clear to me to allow NFTs. (But note the absence of any discussion of derivative works...)

Re: DALL·E now available in beta

#318

Post by berberous »

napier wrote: post #263
berberous wrote: post #20

Earlier quoted context omitted.

I think they are reacting to competition. MidJourney is amazing, was easier to get into, gives you commercial rights, and frankly I found more fun to use and even better output in most instances.

The only thing I don’t like about MidJourney is the Discord based interface. I think I can grok why Dave chose this route as it bakes in an active community element and allows users to pick up prompt engineering techniques osmotically… but I’d prefer a clean DALL-E style app and cli / api access.

In case you don’t know, you can at least PM the MidJourney bot so you have an uncluttered workspace.

It’s clearly personally preference, but I loathe Discord but love it for MidJourney. As you said, there’s an interactive element where I see other people doing cool things and adapting part of their prompts and vice versa. It really is fun. And when you do it in a PM, you have all your efforts saved. DALL-E is pretty clunky in that you have to manually save an image or lose it once your history rolls off.

Re: DALL·E now available in beta

#319

Post by ZeWaka »

eigenvalue wrote: post #302
thorum wrote: post #182

Earlier quoted context omitted.

DALLE images are still only 1024 px wide. Which has its uses, but I don’t think the stock photo industry is in real danger until someone figures out a better AI superresolution system that can produce larger and more detailed images.

I've been using this app to upscale the images to 4000x4000, and it works amazingly well (there is also a version for Android): https://apps.apple.com/us/app/waifu2x/id1286485858 I paid extra to get the higher quality model using the in-app purchase option. It crushes the phone's battery life, but runs in only ~10 seconds on an iPhone 13 Pro for a single 1000x1000 input image.

I mean, waifu2x and similar waifuxx libraries are free and open-source, there's really no reason to pay for it if you're working on a desktop.

Re: DALL·E now available in beta

#320

Post by cmarschner »

jowday wrote: post #306
isoprophlex wrote: post #296

Wait until someone trains a model like this, for porn. There seems to be a post-DALLE obscenity detector on openAI's tool, as so far I've found it to be entirely robust against deliberate typos designed to avoid simple 'bad word lists'. Ask it for a "pruple violon" and you get purple violins... you get the deal. "Metastable" prompts that may or may not generate obscene (content with nudity, guns, violence as I've fou…

If I had to guess, I'd bet they have a supervised classifier trained to recognize bad content (violence, porn, etc) that they use to filter the generated images before passing them to the user, on top of the bad word lists.

Most likely they just take the one from bing. Or, if they trained a better one, it goes vice versa sooner or later
Post reply on HN